GeorgiosIoannouCoder
commited on
Commit
•
c7742ac
1
Parent(s):
89ada58
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,325 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#############################################################################################################################
|
2 |
+
# Filename : app.py
|
3 |
+
# Description: A Streamlit application to utilize five models back to back
|
4 |
+
# Models used:
|
5 |
+
# 1. Visual Question Answering (VQA).
|
6 |
+
# 2. Fill-Mask.
|
7 |
+
# 3. Text2text Generation.
|
8 |
+
# 4. Text Generation.
|
9 |
+
# 5. Topic.
|
10 |
+
# Author : Georgios Ioannou
|
11 |
+
#
|
12 |
+
# Copyright © 2024 by Georgios Ioannou
|
13 |
+
#############################################################################################################################
|
14 |
+
|
15 |
+
# Import libraries.
|
16 |
+
|
17 |
+
import streamlit as st # Build the GUI of the application.
|
18 |
+
import torch # Load Salesforce/blip model(s) on GPU.
|
19 |
+
|
20 |
+
from bertopic import BERTopic # Topic model inference.
|
21 |
+
from PIL import Image # Open and identify a given image file.
|
22 |
+
from transformers import (
|
23 |
+
pipeline,
|
24 |
+
BlipProcessor,
|
25 |
+
BlipForQuestionAnswering,
|
26 |
+
) # VQA model inference.
|
27 |
+
|
28 |
+
#############################################################################################################################
|
29 |
+
|
30 |
+
# Function to apply local CSS.
|
31 |
+
|
32 |
+
|
33 |
+
def local_css(file_name):
|
34 |
+
with open(file_name) as f:
|
35 |
+
st.markdown(f"<style>{f.read()}</style>", unsafe_allow_html=True)
|
36 |
+
|
37 |
+
|
38 |
+
#############################################################################################################################
|
39 |
+
|
40 |
+
# Model 1.
|
41 |
+
# Model 1 gets input from the user.
|
42 |
+
# User -> Model 1
|
43 |
+
|
44 |
+
# Load the Visual Question Answering (VQA) model directly.
|
45 |
+
# Using transformers.
|
46 |
+
|
47 |
+
|
48 |
+
@st.cache_resource
|
49 |
+
def load_model_blip():
|
50 |
+
blip_processor_base = BlipProcessor.from_pretrained("Salesforce/blip-vqa-base")
|
51 |
+
blip_model_base = BlipForQuestionAnswering.from_pretrained(
|
52 |
+
"Salesforce/blip-vqa-base"
|
53 |
+
)
|
54 |
+
|
55 |
+
# Backup model.
|
56 |
+
# blip_processor_large = BlipProcessor.from_pretrained("Salesforce/blip-vqa-capfilt-large")
|
57 |
+
# blip_model_large = BlipForQuestionAnswering.from_pretrained("Salesforce/blip-vqa-capfilt-large")
|
58 |
+
# return blip_processor_large, blip_model_large
|
59 |
+
|
60 |
+
return blip_processor_base, blip_model_base
|
61 |
+
|
62 |
+
|
63 |
+
# General function for any Salesforce/blip model(s).
|
64 |
+
# VQA model.
|
65 |
+
|
66 |
+
|
67 |
+
def generate_answer_blip(processor, model, image, question):
|
68 |
+
# Prepare image + question.
|
69 |
+
|
70 |
+
inputs = processor(images=image, text=question, return_tensors="pt")
|
71 |
+
|
72 |
+
generated_ids = model.generate(**inputs, max_length=50)
|
73 |
+
|
74 |
+
generated_answer = processor.batch_decode(generated_ids, skip_special_tokens=True)
|
75 |
+
|
76 |
+
return generated_answer
|
77 |
+
|
78 |
+
|
79 |
+
# Generate answer from the Salesforce/blip model(s).
|
80 |
+
# VQA model.
|
81 |
+
|
82 |
+
|
83 |
+
@st.cache_resource
|
84 |
+
def generate_answer(image, question):
|
85 |
+
answer_blip_base = generate_answer_blip(
|
86 |
+
processor=blip_processor_base,
|
87 |
+
model=blip_model_base,
|
88 |
+
image=image,
|
89 |
+
question=question,
|
90 |
+
)
|
91 |
+
|
92 |
+
# answer_blip_large = generate_answer_blip(blip_processor_large, blip_model_large, image, question)
|
93 |
+
# return answer_blip_large
|
94 |
+
|
95 |
+
return answer_blip_base
|
96 |
+
|
97 |
+
|
98 |
+
#############################################################################################################################
|
99 |
+
|
100 |
+
# Model 2.
|
101 |
+
# Model 2 gets input from Model 1.
|
102 |
+
# User -> Model 1 -> Model 2
|
103 |
+
|
104 |
+
|
105 |
+
@st.cache_resource
|
106 |
+
def load_model_fill_mask():
|
107 |
+
return pipeline(task="fill-mask", model="bert-base-uncased")
|
108 |
+
|
109 |
+
|
110 |
+
#############################################################################################################################
|
111 |
+
|
112 |
+
# Model 3.
|
113 |
+
# Model 3 gets input from Model 2.
|
114 |
+
# User -> Model 1 -> Model 2 -> Model 3
|
115 |
+
|
116 |
+
|
117 |
+
@st.cache_resource
|
118 |
+
def load_model_text2text_generation():
|
119 |
+
return pipeline(
|
120 |
+
task="text2text-generation", model="facebook/blenderbot-400M-distill"
|
121 |
+
)
|
122 |
+
|
123 |
+
|
124 |
+
#############################################################################################################################
|
125 |
+
|
126 |
+
# Model 4.
|
127 |
+
# Model 4 gets input from Model 3.
|
128 |
+
# User -> Model 1 -> Model 2 -> Model 3 -> Model 4
|
129 |
+
|
130 |
+
|
131 |
+
@st.cache_resource
|
132 |
+
def load_model_fill_text_generation():
|
133 |
+
return pipeline(task="text-generation", model="gpt2")
|
134 |
+
|
135 |
+
|
136 |
+
#############################################################################################################################
|
137 |
+
|
138 |
+
# Model 5.
|
139 |
+
# Model 5 gets input from Model 4.
|
140 |
+
# User -> Model 1 -> Model 2 -> Model 3 -> Model 4 -> Model 5
|
141 |
+
|
142 |
+
|
143 |
+
@st.cache_resource
|
144 |
+
def load_model_bertopic1():
|
145 |
+
return BERTopic.load(path="davanstrien/chat_topics")
|
146 |
+
|
147 |
+
|
148 |
+
@st.cache_resource
|
149 |
+
def load_model_bertopic2():
|
150 |
+
return BERTopic.load(path="MaartenGr/BERTopic_ArXiv")
|
151 |
+
|
152 |
+
|
153 |
+
#############################################################################################################################
|
154 |
+
# Page title and favicon.
|
155 |
+
|
156 |
+
st.set_page_config(page_title="Visual Question Answering", page_icon="❓")
|
157 |
+
|
158 |
+
#############################################################################################################################
|
159 |
+
|
160 |
+
# Load the Salesforce/blip model directly.
|
161 |
+
|
162 |
+
if torch.cuda.is_available():
|
163 |
+
device = torch.device("cuda")
|
164 |
+
# elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
|
165 |
+
# device = torch.device("mps")
|
166 |
+
else:
|
167 |
+
device = torch.device("cpu")
|
168 |
+
|
169 |
+
blip_processor_base, blip_model_base = load_model_blip()
|
170 |
+
blip_model_base.to(device)
|
171 |
+
|
172 |
+
#############################################################################################################################
|
173 |
+
# Main function to create the Streamlit web application.
|
174 |
+
#
|
175 |
+
# 5 MODEL INFERENCES.
|
176 |
+
# User Input = Image + Question About The Image.
|
177 |
+
# User -> Model 1 -> Model 2 -> Model 3 -> Model 4 -> Model 5
|
178 |
+
|
179 |
+
|
180 |
+
def main():
|
181 |
+
try:
|
182 |
+
#####################################################################################################################
|
183 |
+
|
184 |
+
# Load CSS.
|
185 |
+
|
186 |
+
local_css("styles/style.css")
|
187 |
+
|
188 |
+
#####################################################################################################################
|
189 |
+
|
190 |
+
# Title.
|
191 |
+
|
192 |
+
title = f"""<h1 align="center" style="font-family: monospace; font-size: 2.1rem; margin-top: -4rem">
|
193 |
+
Georgios Ioannou's Visual Question Answering</h1>"""
|
194 |
+
st.markdown(title, unsafe_allow_html=True)
|
195 |
+
# st.title("ChefBot - Automated Recipe Assistant")
|
196 |
+
|
197 |
+
#####################################################################################################################
|
198 |
+
|
199 |
+
# Subtitle.
|
200 |
+
|
201 |
+
subtitle = f"""<h2 align="center" style="font-family: monospace; font-size: 1.5rem; margin-top: -2rem">
|
202 |
+
CUNY Tech Prep Tutorial 4</h2>"""
|
203 |
+
st.markdown(subtitle, unsafe_allow_html=True)
|
204 |
+
|
205 |
+
#####################################################################################################################
|
206 |
+
|
207 |
+
# Image.
|
208 |
+
|
209 |
+
image = "./ctp.png"
|
210 |
+
left_co, cent_co, last_co = st.columns(3)
|
211 |
+
with cent_co:
|
212 |
+
st.image(image=image)
|
213 |
+
|
214 |
+
#####################################################################################################################
|
215 |
+
|
216 |
+
# User input (Image).
|
217 |
+
image = st.file_uploader("Choose an image...", type=["jpg", "jpeg", "png"])
|
218 |
+
|
219 |
+
if image is not None:
|
220 |
+
bytes_data = image.getvalue()
|
221 |
+
|
222 |
+
with open(image.name, "wb") as file:
|
223 |
+
|
224 |
+
file.write(bytes_data)
|
225 |
+
st.image(image, caption="Uploaded Image.", use_column_width=True)
|
226 |
+
raw_image = Image.open(image.name).convert("RGB")
|
227 |
+
|
228 |
+
# User input (Question).
|
229 |
+
question = st.text_input("What's your question?")
|
230 |
+
|
231 |
+
#############################################################################################################
|
232 |
+
|
233 |
+
if question != "":
|
234 |
+
# Model 1.
|
235 |
+
with st.spinner(
|
236 |
+
text="VQA inference..."
|
237 |
+
): # Spinner to keep the application interactive.
|
238 |
+
# Model inference.
|
239 |
+
|
240 |
+
answer = generate_answer(raw_image, question)[0]
|
241 |
+
st.success(f"VQA: {answer}")
|
242 |
+
|
243 |
+
bbu_pipeline = load_model_fill_mask()
|
244 |
+
text = (
|
245 |
+
"I love " + answer + " and I would like to know how to [MASK]."
|
246 |
+
)
|
247 |
+
|
248 |
+
#########################################################################################################
|
249 |
+
|
250 |
+
# Model 2.
|
251 |
+
with st.spinner(
|
252 |
+
text="Fill-Mask inference..."
|
253 |
+
): # Spinner to keep the application interactive.
|
254 |
+
# Model inference.
|
255 |
+
bbu_pipeline_output = bbu_pipeline(text)
|
256 |
+
bbu_output = bbu_pipeline_output[0]["sequence"]
|
257 |
+
st.success(f"Fill-Mask: {bbu_output}")
|
258 |
+
|
259 |
+
facebook_pipeline = load_model_text2text_generation()
|
260 |
+
utterance = bbu_output
|
261 |
+
|
262 |
+
#########################################################################################################
|
263 |
+
|
264 |
+
# Model 3.
|
265 |
+
with st.spinner(
|
266 |
+
text="Text2text Generation inference..."
|
267 |
+
): # Spinner to keep the application interactive.
|
268 |
+
# Model inference.
|
269 |
+
facebook_pipeline_output = facebook_pipeline(utterance)
|
270 |
+
facebook_output = facebook_pipeline_output[0]["generated_text"]
|
271 |
+
st.success(f"Text2text Generation: {facebook_output}")
|
272 |
+
|
273 |
+
gpt2_pipeline = load_model_fill_text_generation()
|
274 |
+
|
275 |
+
#########################################################################################################
|
276 |
+
|
277 |
+
# Model 4.
|
278 |
+
with st.spinner(
|
279 |
+
text="Fill Text Generation inference..."
|
280 |
+
): # Spinner to keep the application interactive.
|
281 |
+
# Model inference.
|
282 |
+
gpt2_pipeline_output = gpt2_pipeline(facebook_output)
|
283 |
+
gpt2_output = gpt2_pipeline_output[0]["generated_text"]
|
284 |
+
st.success(f"Fill Text Generation: {gpt2_output}")
|
285 |
+
|
286 |
+
#########################################################################################################
|
287 |
+
|
288 |
+
# Model 5.
|
289 |
+
topic_model_1 = load_model_bertopic1()
|
290 |
+
topic, prob = topic_model_1.transform(gpt2_pipeline_output)
|
291 |
+
topic_model_1_output = topic_model_1.get_topic_info(topic[0])[
|
292 |
+
"Representation"
|
293 |
+
][0]
|
294 |
+
st.success(
|
295 |
+
f"Topic(s) from davanstrien/chat_topics: {topic_model_1_output}"
|
296 |
+
)
|
297 |
+
|
298 |
+
topic_model_2 = load_model_bertopic2()
|
299 |
+
topic, prob = topic_model_2.transform(gpt2_pipeline_output)
|
300 |
+
topic_model_2_output = topic_model_2.get_topic_info(topic[0])[
|
301 |
+
"Representation"
|
302 |
+
][0]
|
303 |
+
st.success(
|
304 |
+
f"Topic(s) from MaartenGr/BERTopic_ArXiv: {topic_model_1_output}"
|
305 |
+
)
|
306 |
+
except Exception as e:
|
307 |
+
# General exception/error handling.
|
308 |
+
|
309 |
+
st.error(e)
|
310 |
+
|
311 |
+
# GitHub repository of author.
|
312 |
+
|
313 |
+
st.markdown(
|
314 |
+
f"""
|
315 |
+
<p align="center" style="font-family: monospace; color: #FAF9F6; font-size: 1rem;"><b> Check out our
|
316 |
+
<a href="https://github.com/GeorgiosIoannouCoder/" style="color: #FAF9F6;"> GitHub repository</a></b>
|
317 |
+
</p>
|
318 |
+
""",
|
319 |
+
unsafe_allow_html=True,
|
320 |
+
)
|
321 |
+
|
322 |
+
|
323 |
+
#############################################################################################################################
|
324 |
+
if __name__ == "__main__":
|
325 |
+
main()
|